Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
Abstract
Future Chunk Prediction improves autoregressive video diffusion by enforcing future-chunk supervision, yielding more physically plausible video generation with reduced training cost.
Autoregressive video diffusion models generate videos chunk by chunk, yet producing plausible continuations does not guarantee that they capture how scenes evolve. Under teacher forcing, strong temporal redundancy may allow local visual continuity to serve as a shortcut for current-chunk denoising, reducing the pressure to model visual evolution. We introduce Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with teacher-forced denoising that predicts temporally separated chunks without access to the intervening visual context. Drawing supervision directly from training videos, FCP requires no additional pretrained visual encoder and leaves autoregressive inference unchanged. Our method achieves superior physical plausibility over standard teacher-forced autoregressive training with a lower overall training compute budget, highlighting the effectiveness and efficiency of future-chunk supervision for modeling visual evolution. Intriguingly, our visualizations suggest that FCP fosters a more structured representation space, which may be valuable for building autoregressive world models.
Community
🔭 Introducing Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation — teaching autoregressive video models to learn how the world evolves and better anticipate future dynamics through generative future prediction.
Our method delivers:
🌏 Better physical plausibility: +11.2% (I2V) and +9.4% (V2V) on Physics-IQ.
⚡ Higher training efficiency: 57.9% less compute and 63.4% less data.
Learn more:
💻 GitHub: https://github.com/TencentARC/Looking-Ahead
🌐 Project page: https://hellorpg.github.io/looking-ahead/
NOTE on Training Data & Future Improvements
Our current training data still primarily consists of general-purpose scenarios, with limited targeted coverage of diverse physical phenomena. We are actively exploring ways to incorporate more diverse and physics-related data into future training, aiming to continuously improve the model's capabilities and make it more competitive with specialized models.
Get this paper in your agent:
hf papers read 2610.29989 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper